Why We Stopped Buying Domain Authority and Computed Our Own
A rented quality metric makes every qualification decision unexplainable and every renewal a dependency. Here is the signal set we replaced it with, how the weights were chosen, and what it costs to be wrong.
Every link-building pipeline needs a gate. Before an agent writes copy for a publisher, uploads a logo, and files a form, something has to decide whether that publisher is worth the attempt at all. For most of the category, that gate is a number bought from a vendor — a single composite score that arrives with no derivation and no recourse.
We built the first version of LinkForge that way too. It worked until an operator asked a reasonable question about a rejected publisher: why? We could point at a threshold. We could not explain the number the threshold was applied to.
What a rented score actually costs
Three costs, in ascending order of severity. The metered API bill is the obvious one and the least important. The second is coupling: a qualification stage that cannot run without a third-party quota is a pipeline that stops when someone forgets to renew a key. The third is the one that matters — an opaque score cannot be argued with, so it cannot be improved.
If you cannot show an operator why a publisher scored 47, you have not built a qualification system. You have bought a coin flip with a decimal point.
The signal set
The LinkForge Authority Score is assembled from eight public signals, each independently observable and each stored with its own contribution to the total. Nothing in the list requires a paid data partner.
Hard-fail checks sit in front of the score, not inside it. A publisher that is unreachable, de-indexed, serving over plain HTTP, or classified into an excluded vertical never reaches the weighting step. Scoring a site we would refuse to submit to is wasted compute and a misleading record.
Where the weights came from, and why they are provisional
The initial weights are a judgment call by the engineers and SEO operators who ran these campaigns by hand for years. We are explicit about that in the product: early scores are provisional. The weights only become defensible once enough placements have completed a full lifecycle for us to regress against real outcomes — acceptance rate and, more importantly, whether the link is still live ninety days later.
Because each signal's contribution is persisted per publisher, recalibration is a re-weighting exercise rather than a re-crawl. That property is the whole argument for computing the score yourself.
The middle band belongs to a human
Scores in the 40 to 59 range, and any otherwise-qualified publisher where the model's confidence is low, route to an explicit operator review queue. This was not the original design. The original design auto-rejected them, which quietly discarded a meaningful slice of usable inventory and gave us no signal about the boundary.
Treating manual review as a first-class state rather than a stall changed the qualification state machine and, unexpectedly, became the fastest source of calibration data we have.